1052 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO.
2, FEBRUARY 2021
Joint Transceiver and Large Intelligent Surface
Design for Massive MIMO mmWave Systems
Peilan Wang , Jun Fang , Senior Member, IEEE, Linglong Dai , Senior Member, IEEE,
and Hongbin Li, Fellow, IEEE
Abstract— Large intelligent surface (LIS) has recently emerged
as a potential low-cost solution to reshape the wireless propaga-
tion environment for improving the spectral efficiency. In this
article, we consider a downlink millimeter-wave (mmWave)
multiple-input-multiple-output (MIMO) system, where an LIS
is deployed to assist the downlink data transmission from a base
station (BS) to a user equipment (UE). Both the BS and the UE
are equipped with a large number of antennas, and a hybrid
analog/digital precoding/combining structure is used to reduce
the hardware cost and energy consumption. We aim to maximize
the spectral efficiency by jointly optimizing the LIS’s reflection
coefficients and the hybrid precoder (combiner) at the BS (UE).
To tackle this non-convex problem, we reformulate the complex
optimization problem into a much more friendly optimization
problem by exploiting the inherent structure of the effective (cas-
cade) mmWave channel. A manifold optimization (MO)-based
algorithm is then developed. Simulation results show that by
carefully devising LIS’s reflection coefficients, our proposed
method can help realize a favorable propagation environment
with a small channel matrix condition number. Besides, it can
achieve a performance comparable to those of state-of-the-art
algorithms, while at a much lower computational complexity.
Index Terms— mmWave communications, large intelligent sur-
face (LIS), joint transceiver-LIS design.
I. I NTRODUCTION
M ILLIMETER-WAVE (mmWave) communication is
regarded as a promising technology for future cellular
networks due to its large available bandwidth and the potential
to offer gigabits-per-second communication data rates [1]–[4].
Utilizing large antenna arrays is essential for mmWave systems
as mmWave signals incur a much higher free-space path
Manuscript received May 19, 2020; revised August 11, 2020 and
October 9, 2020; accepted October 9, 2020. Date of publication October 20,
2020; date of current version February 11, 2021. This work was supported in
part by the National Science Foundation of China under Grant 61829103 and
Grant 61871091, and in part by the Natural Science Foundation of China
for Outstanding Young Scholars under Grant 61722109. The associate editor
coordinating the review of this article and approving it for publication was
C.-K. Wen. (Corresponding author: Jun Fang.)
Peilan Wang and Jun Fang are with the National Key Labora-
tory of Science and Technology on Communications, University of
Electronic Science and Technology of China, Chengdu 611731, China
(e-mail: peilan_wangle@std.uestc.edu.cn; junfang@uestc.edu.cn).
Linglong Dai is with the Department of Electronic Engineering, Tsinghua
University, Beijing 100084, China (e-mail: daill@tsinghua.edu.cn).
Hongbin Li is with the Department of Electrical and Computer Engi-
neering, Stevens Institute of Technology, Hoboken, NJ 07030 USA (e-mail:
hongbin.li@stevens.edu).
Color versions of one or more of the figures in this article are available
online at https://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TWC.2020.3030570
1536-1276 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
loss compared to microwave signals below 6 GHz [5], [6].
Large antenna arrays can help form directional beams to
compensate for severe path loss incurred by mmWave signals.
On the other hand, high directivity makes mmWave communi-
cations vulnerable to blockage events, which can be frequent in
indoor and dense urban environments. For instance, due to the
narrow beamwidth of mmWave signals, a very small obstacle,
such as a person’s arm, can effectively block the link [7], [8].
To address this blockage issue, large intelligent surface has
been recently introduced to improve the spectral efficiency
and coverage of mmWave systems [9]–[12]. Large intelligent
surface (LIS), also referred to intelligent reflecting surface
(IRS), has emerged as a promising technology to realize
a smart and programmable wireless propagation environ-
ment via software-controlled reflection [13]–[18]. Specifically,
LIS is a two-dimensional artificial structure, consisting of a
large number of low-cost, passive, reconfigurable reflecting
elements. With the help of a smart controller, the inci-
dent signal can be reflected with reconfigurable phase shifts
and amplitudes. By properly adjusting the phase shifts, the
LIS-assisted communications system can achieve desired prop-
erties, such as extending signal coverage [19], improving
energy efficiency [20], mitigating interference [21], [22],
enhancing system security [23], and so on.
A key problem of interest in LIS-assisted communication
systems is to jointly devise the reflection coefficients at the
LIS and the active precoding matrix at the BS to optimize
the system performance. A plethora of studies have been
conducted to investigate the problem under different system
setups and assumptions, e.g. [9], [11], [19]–[28]. Among
them, most focused on single-input-single-output (SISO) or
multiple-input-single-output (MISO) systems [9], [19]–[23].
For the MIMO scenario where both the BS and the UE are
equipped with multiple antennas, [24] proposed to optimize
the spectral efficiency via maximizing the Frobenius-norm of
the effective channel from the BS to the UE. Nevertheless,
maximizing the Frobenius-norm of the effective channel usu-
ally results in a large condition number, and as a result, its
performance improvement is limited. In [25], an alternating
optimization (AO)-based method was developed to maximize
the capacity of LIS-aided point-to-point MIMO systems via
jointly optimizing LIS’s reflection coefficients and the MIMO
transmit covariance matrix. Specifically, the AO-based method
sequentially and iteratively optimizes one reflection coefficient
at a time by fixing the other reflection coefficients. Although
achieving superior performance, this sequential optimization
WANG et al.: JOINT TRANSCEIVER AND LIS DESIGN FOR MASSIVE MIMO mmWAVE SYSTEMS
Fig. 1. LIS-assisted mmWave hybrid downlink system.
incurs an excessively high computational complexity for
practical systems. In [11], two optimization schemes were
proposed to enhance the channel capacity for LIS-assisted
mmWave MIMO systems. Nevertheless, this work is confined
to the scenario where only a single data stream is transmitted
from the BS to the UE. To improve the bit error rate (BER)
performance, [26] proposed a geometric mean decomposition
(GMD)-based beamforming scheme for LIS-assisted mmWave
MIMO systems. To simplify the problem, only the angle of
arrival (AoA) of the line-of-sight (LOS) BS-LIS link and the
angle of departure (AoD) of the LOS LIS-UE path were uti-
lized to design the reflection coefficients. This simplification,
however, comes at the cost of sacrificing the spectral efficiency.
Besides the above works, other studies [27], [28] considered
multiple-user (MU)-MIMO scenarios. For example, [27] con-
sidered using the LIS at the cell boundary to assist downlink
transmission to cell-edge users.
In this article, we consider a point-to-point LIS-assisted
mmWave systems with large-scale antenna arrays at the trans-
mitter and receiver, where an LIS is deployed to assist the
downlink transmission from the transmitter (i.e. BS) to the
receiver (i.e. UE). Our objective is to maximize the spectral
efficiency of the LIS-assisted mmWave system by jointly
optimizing the LIS reflection coefficients and the hybrid pre-
coding/combining matrices associated with the BS and the UE.
To address this non-convex optimization problem, we first
decouple the LIS design from hybrid precoder/combiner
design. We then focus on the reflection coefficient optimization
problem. By exploiting the inherent structure of the effec-
tive mmWave channel, it is shown that this complex LIS
optimization problem can be reformulated into an amiable
problem that maximizes the spectral efficiency with respect
to the passive beamforming gains (which have an explicit
expression of the reflection coefficients) associated with the
BS-LIS-UE composite paths. A manifold-based optimization
method is then developed to solve the LIS (also referred to
as passive beamforming) design problem. Simulation results
show that our proposed method can help create a favorable
Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
1053
propagation environment with a small channel matrix con-
dition number. Also, the proposed method exhibits a sig-
nificant performance improvement over the sum-path-gain
maximization method [24] and achieves a performance similar
to that of the AO-based method [25], while at a much lower
computational complexity.
The rest of the paper is organized as follows. In Section II,
the system model and the joint active and passive beamforming
problem are discussed. The passive beamforming problem is
studied in Section III, where a manifold-based optimization
method is developed. The hybrid precoding/combining design
problem is considered in Section IV. The comparison with the
state-of-art algorithms is discussed in Section V. Simulation
results are provided in Section VI, followed by concluding
remarks in Section VII.
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
A. System Model
We consider a point-to-point mmWave MIMO system as
illustrated in Fig. 1, where the BS and the UE are equipped
with a large number of antennas, and the LIS is deployed to
assist the downlink data transmission from the BS to the UE.
For simplicity, we assume that the direct link between the BS
and the UE is blocked due to unfavorable propagation con-
ditions. Nevertheless, our proposed scheme can be extended
to the scenario in which a direct link between the BS and
the UE is available. The BS is equipped with Nt antennas
and Rt radio frequency (RF) chains, and the UE is equipped
with Nr antennas and Rr RF chains, where Rt  Nt and
Rr  Nr . To exploit the channel diversity, multiple, say Ns ,
data streams are simultaneously sent from the BS to the UE,
where Ns ≤ min{Rt , Rr }. At the BS, a digital baseband
precoder F BB ∈ CRt ×Ns is first applied to the transmitted
signal s ∈ CNs , then followed by an analog RF beamformer
F RF ∈ CNt ×Rt . The transmitted signal can be written as
x = F RF F BB s, (1)
1054
where s is assumed to satisfy E[ssH ] = I. Also, a transmit
power constraint is placed on the hybrid precoding matrix
F RF F BB , i.e. F RF F BB 2F ≤ ρ, ρ is the transmit power.
The transmitted signal arrives at the UE via propagating
through the BS-LIS-UE channel, where the LIS comprises M
passive reflecting elements and each element behaves like a
single physical point which combines all the received signals
and then re-scatters the combined signal with an adjustable
phase shift [21]. Let G ∈ CM×Nt denote the channel from the
BS to the LIS, and R ∈ CNr ×M denote the LIS-UE channel.
Also, let φm ∈ [0, 2π] denote the phase shift introduced by
the mth element of the LIS, and
Φ  diag(ejφ1 , · · · , ejφM ) ∈ CM×M .
The signal received by the UE can thus be given as
y = WH
BB W RF H eff F RF F BB s + W BB W RF n,
H H H
Nr ×Rr Rr ×Ns
where W RF ∈ C and W BB ∈ C represent the
analog combiner and the digital baseband combiner, respec-
tively, H eff  RΦG stands for the effective (i.e. cascade)
channel from the BS to the UE, and n ∈ CNr ∼ CN (0, σ 2 I)
denotes the additive white Gaussian noise.
In this article, we assume that perfect channel state infor-
mation (CSI) is known for joint transceiver and LIS design.
Channel estimation for LIS-assisted systems is an important
and challenging issue that has been studied in a variety of
works, e.g. [29]–[32]. In particular, [30]–[32] discussed how
to estimate the channel for LIS-assisted mmWave systems.
Suppose that the transmitted signal follows a Gaussian dis-
tribution. The achievable spectral efficiency can be expressed
as [33]

1 |F RF (i, j)| = |W RF (i, j)| = 1, ∀i, j,
R = log2 det I + 2 (W RF W BB )† H eff F RF F BB
σ H eff = RΦG

×F BB F RF H eff (W RF W BB ) . (4)
H H H
where A† denotes the Moore-Penrose pseudo inverse of the
matrix A. It should be noted that since the analog precoder and
combiner are implemented by analog phase shifters, entries of
F RF and W RF have constant modulus.
B. Channel Model
Due to the small wavelength, mmWave has limited
ability to diffract around obstacles. As a result, mmWave
channels exhibit a sparse multipath structure and are usually
characterized by the Saleh-Valenzuela (SV) channel
model [26], [34]–[37]. Suppose uniform linear arrays (ULAs)
are employed at the BS and the UE, and the LIS is a
uniform planar array (UPA) consisting of a large number of
reconfigurable passive elements. The BS-LIS channel G and
the LIS-UE channel R can be given as
 Φ = diag(ejφ1 , · · · , ejφM ).
Nt M 
P
G= α̃i aRLIS (θir , ηir )aH t
TBS (γi ), (5)
P i=1
 to search for a hybrid precoding/combining matrix to approx-
M Nr 
L
R= β̃i aRUE (γir )aH t t
TLIS (θi , ηi ), (6)
L i=1
Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 2, FEBRUARY 2021
where P (L) is the total number of signal paths between the
BS and the LIS (the LIS and the UE), θir (θit ) and ηir (ηit )
denote the azimuth and elevation angles of arrival (departure)
associated with the LIS, γir (γit ) represents the angle of
arrival (departure) associated with the UE (BS), α̃i (β̃i ) is
the complex channel gain, aRj , j ∈ {LIS, UE} and aTi , i ∈
{BS, LIS} denote the normalized array response vectors asso-
ciated with the receiver and the transmitter, respectively.
Specifically, for ULA with N antennas, the normalized array
response vector is given by
1 2πd 2πd
a(γ) = √ [1 · · · ej λ (n−1) sin(γ) · · · ej λ (N −1) sin(γ) ]T ,
N
(2) (7)
where d and λ are the antenna spacing and the signal wave-
length. For UPA with M = My ×Mz elements, the normalized
(3) array response vector can be written as
1 2πd
a(θ, η) = √ [1 · · · ej λ ((m1 −1) cos(η) sin(θ)+(m2 −1) sin(η))
M
2πd
· · · ej λ ((My −1) cos(η) sin(θ)+(Mz −1) sin(η)) ]T . (8)
C. Problem Formulation
Our objective is to jointly devise the hybrid precod-
ing/combining matrices and the passive beamforming matrix
Φ to maximize the spectral efficiency (4):
 1
max log2 det I + 2 (W RF W BB )† H eff
{F RF ,F BB ,W RF ,W BB ,Φ} σ

×F RF F BB F HBB F RF H eff (W RF W BB )
H H
s.t. F RF F BB 2F ≤ ρ,
Φ = diag(ejφ1 , · · · , ejφM ). (9)
Such an optimization problem is challenging due to the non-
convexity of the objective function and the per-element unit-
modulus constraint placed on analog precoding and combining
matrices.
To simplify the problem, we first ignore the constraint
introduced by the hybrid analog/digital structure and consider
a fully digital precoder/combiner. Let F ∈ CNt ×Ns be a fully
digital precoder which has the same size as the hybrid pre-
coding matrix F RF F BB , and W ∈ CNr ×Ns be a fully digital
combiner which has the same size as the hybrid combining
matrix W RF W BB . The problem (9) can be simplified as
 1 
max log2 det I + 2 (W )† H eff F F H H H
eff W
{F ,W ,Φ} σ
s.t. F 2F ≤ ρ
H eff = RΦG
(10)
Once an optimal fully digital precoder/combiner is obtained,
we can employ the manifold optimization-based method [38]
imate the optimal fully digital precoder/combiner. Such a
strategy can be well justified because it has been shown
WANG et al.: JOINT TRANSCEIVER AND LIS DESIGN FOR MASSIVE MIMO mmWAVE SYSTEMS 1055

in many previous studies [33], [35], [36], [38] that, due to values to support multi-stream transmission [39]. To address
the sparse scattering nature of mmWave channels, hybrid this issue, we will develop a new approach to solve (14). Our
beamforming/combining with a small number of RF chains proposed method obtains more balanced singular values of Σ1
can asymptotically approach the performance of fully digital which can help substantially improve the spectral efficiency,
beamforming/combining for a sufficiently large number of as suggested by our empirical results.
antennas at the transceiver.
Given a fixed Φ (i.e. H eff ), the optimal F and W can be III. P ROPOSED PASSIVE B EAMFORMING D ESIGN M ETHOD
obtained via the singular value decomposition (SVD) of H eff .
Optimizing (14) is much more challenging than directly
Define the effective channel’s ordered SVD as
  H optimizing the Frobenius-norm of H eff , as the singular value
Σ1 0 V1 can not be expressed by H eff in an explicit way. To address
H eff = U ΣV = [U 1 U 2 ]
H
, (11)
0 Σ2 V H 2 this difficulty, we exploit the structure of mmWave chan-
nels to gain insight into the SVD of the effective channel.
where U is an Nr × Q unitary matrix, Σ is a Q × Q diagonal
Recall (5)–(6) and assume |β1 | ≥ |β2 | ≥ .  . . ≥ |βL | and
matrix of singular values, V is an Nt × Q unitary matrix,
in which Q  rank(H eff ). The matrix Σ1 is of dimension |α1 | ≥ |α2 | ≥ · · · ≥ |αP |, where αi  Nt M
P α̃i , and

Ns × Ns and V 1 is of dimension Nt × Ns . Then with a βi  L β̃i . We will justify the order of {βl } and {αp }
MNr
fixed Φ, the optimal solution to (10) is given as later. The effective channel H eff = RΦG can be written as
F opt = V 1 Λ1/2 , W opt = U 1 , (12) RΦG

L 
where Λ = diag(p1 , . . . , pNs ), pi = max(1/λ −
= βi aRUE (γi )aTLIS (θi , ηi ) Φ
r H t t
σ 2 /Σ21 (i, i), 0), i = 1, . . . , Ns denotes the optimal amount
i=1
of power allocated
Nto the ith data stream, 1/λ is the water 
P 
s
level satisfying i=1 i = ρ. Thanks to the massive array
p × αj aRLIS (θjr , ηjr )aH (γ t
)
TBS j
gain provided by the LIS and large number of antennas at the j=1
BS, the effective signal-to-noise ratio (SNR) is large, in which

L 
P
case an equal power allocation scheme is near-optimal [35]. βi αj aRUE (γir )aHTLIS (θi , ηi )ΦaRLIS (θj , ηj )aTBS(γj )
t t r r H t
=
Therefore we can approximate F opt as: i=1 j=1
  
 dij
ρ
F opt ≈ V 1. (13) 
L 
P
Ns = βi αj dij aRUE (γir )aH t
TBS (γj )
Substituting the optimal fully digital precoder and combiner i=1 j=1

into (10), we arrive at a problem which concerns only the = ARUE DAH
TBS , (17)
optimization of the passive beamforming matrix Φ:
  where D is an L × P matrix with its (i, j)th entry given by
ρ 2 D(i, j) = βi αj dij , and dij is defined as
max log2 det I + Σ
Φ Ns σ 2 1
dij  aH t t r r
TLIS (θi , ηi )ΦaRLIS (θj , ηj )
s.t. Φ = diag(e , · · · , ejφM ).
jφ1
(14)
= v H (a∗TLIS (θit , ηit ) ◦ aRLIS (θjr , ηjr ))
The above objective function can be bounded by = v H pij (18)
 
ρ 2
log2 det I + Σ with ◦ stands for the Hadamard (elementwise) product,
Ns σ 2 1 v  diag(ΦH ), and pij  (a∗TLIS (θit , ηit ) ◦ aRLIS (θjr , ηjr )).
(a) ρ Here dij is referred to as the passive beamforming gain
≤ Ns log2 (1 + 2 2 tr(Σ21 ))
Ns σ associated with the (j, i)th BS-LIS-UE composite path which
(b) ρ is composed of the jth path from the BS to the LIS and the
≤ Ns log2 (1 + 2 2 tr(H eff H H eff )), (15) ith path from the LIS to the UE.
Ns σ
Also, ARUE and ATBS in (17) are respectively defined as
where (a) is due to Jensen’s inequality and (b) becomes
equality when Ns = Q. Therefore, alternatively, [24] proposed ARUE  [aRUE (γ1r ) · · · aRUE (γLr )],
to maximize the bound of the spectral efficiency: ATBS  [aTBS (γ1t ) · · · aTBS (γPt )]. (19)
max tr(H eff H H
eff ) For ULA with N antennas, it can be easily verified that
Φ
s.t. Φ = diag(ejφ1 , · · · , ejφM ). (16) |aH (γi )a(γj )| → 0, N → ∞ (20)
Nevertheless, the objective of (16) does not consider the fair- for any sin(γi ) − sin(γj ) = 0. This asymptotic orthogonality
ness among different singular values and thus results in a large also holds valid for other array geometries such as UPA [40].
condition number of H eff . A large condition number indicates Therefore when Nt and Nr are sufficiently large, ARUE and
an unfavorable propagation condition since we prefer the ATBS can be considered as orthonormal matrices in which
power of the channel to be uniformly distributed over singular the columns vectors form an orthonormal set (each column

Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
1056 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 2, FEBRUARY 2021

vector has unit norm and is orthogonal to all the other column technique to address this difficulty. The search space of (23)
vectors). If the phase shift vector v is properly devised such can be regarded as the product of M complex circles, i.e.,
that the off-diagonal elements of D are small relative to entries
on the main diagonal, then H eff = ARUE DAH S × S · · · × S, (24)
TBS can be   
treated as an approximation of the truncated SVD of H eff , M times
in which case the optimization problem (14) turns into
where S  {u ∈ C : uH u = 1} is one complex circle [43].

Ns   The product of such M circles is a submanifold of CM , known
ρ
max log2 1 + |D(i, i)|2 as the complex circle manifold (CCM) [44], which is defined
v Ns σ 2
i=1 as
s.t. D(i, i) = αi βi dii , ∀i ∈ {1, · · · , Ns },
dii = v H pii , M = S M  {u ∈ CM : |ui | = 1, i = 1, 2, · · · , M }. (25)
|dij | = |v H pij | < τ, ∀i = j, More background on optimization on manifolds can be found
v = [ejφ1 · · · ejφM ]H , (21) in [44].
For a smooth real-valued objective function on some mani-
where τ is a small positive value and the constraint |dij | < τ fold, many classical line-search algorithms such as the gradient
is to make sure that D is approximately a non-square diagonal descent method can be applied with certain modifications [45].
matrix such that H eff = ARUE DAH TBS approximates a trun- We start with some basic definitions. The tangent space Tvk M
cated SVD of H eff . One could question that such a constraint is the set consisting of all tangent vectors to M at the point v k
(i.e. |dij | < τ, ∀i = j) not only complicates the problem, but [44]. Here, the tangent space of the complex circle manifold
also confines the solution space of v, and thus may prevent admits a closed-form expression, i.e.,
from achieving a higher spectral efficiency. We will show that
this constraint can be ignored and this omission does not affect Tvk M = {z ∈ CM : {z ◦ (v k )∗ } = 0}, (26)
the validity of our proposed solution. Specifically, we focus on
solving where ◦ denotes the Hadamard product. Since the descent is
performed on the manifold, we need to find the direction of

Ns   the greatest decrease from the tangent space, which is known
ρ
max log2 1 + |D(i, i)|2 as the negative Riemannian gradient. For the complex circle
v Ns σ 2
i=1 manifold, the
Riemannian gradient of the objective function
s.t. D(i, i) = αi βi dii , ∀i ∈ {1, · · · , Ns }, f (v)  − i=1
Ns
log(1 + ai v H P ii v) at the point v k is a
dii = v H pii , tangent vector ∇M f (v k ) given by [44]
v = [ejφ1 · · · ejφM ]H , (22)
∇M f (v k ) = ProjT M (∇f (v
k
))
vk

and we will show that for a sufficiently large M , the solution = ∇f (v ) − {∇f (vk ) ◦ (v k )∗ } ◦ v k , (27)
k

to (22) automatically ensures that off-diagonal entries of D are

small relative to entries on its main diagonal. Thus it is where Proj(·) is the orthogonal projection operator and the
expected that solving (22) leads to an effective solution Euclidean gradient ∇f (v k ) is given as
of (14).

Ns
1 2ai P ii v k
∇f (v ) = −
k
. (28)
i=1
ln 2 1 + ai (v k )H P ii v k
A. Manifold-Based Method for Passive Beamforming Design
In this subsection, we develop a manifold-based method to After we obtain the Riemannian gradient, we can transplant
solve (22). Note that other methods such as the semidefinite many line-search algorithms, such as the gradient descent
relaxation (SDR) method, the block coordinate descent (BCD) method, from the Euclidean spaces to the Riemannian mani-
method [41] and the penalty dual decomposition (PDD) folds [45]. Specifically, we update v k with a step k
method [42] can also be employed to address (22). The reason v̄ k = v k − k ∇M f (v k ), (29)
we choose to use the manifold optimization technique is
that it achieves a good balance between the computational where k is chosen as the Armijo step size [44]. In general, v̄ k
complexity and the performance. Recalling that D(i, i) = does not lie in the complex circle manifold, thus the retraction
αi βi v H pii , the optimization (22) can be rewritten as is needed to map the updated point v̄ k on the tangent space
Tvk M onto the complex circle manifold M with a local

Ns
max log2 (1 + ai v H P ii v) rigidity condition preserving the gradients at v k . In this case,
v the retraction operator is defined as [44]
i=1
s.t. v = [ejφ1 · · · ejφM ]H , (23) 1
v k+1 = R(v̄ k )  v̄ k ◦ . (30)
|v̄ k |
where P ii  pii (pii )H , and ai  Nsρσ2 |αi βi |2 . The main
difficulty involved in solving (23) is the non-convex unit The algorithm above is summarized in Algorithm 1, which is
modulus constraint. We employ the manifold optimization guaranteed to converge to a critical point of (23) [44].

Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: JOINT TRANSCEIVER AND LIS DESIGN FOR MASSIVE MIMO mmWAVE SYSTEMS 1057

Algorithm 1 Proposed Algorithm for Solving (23) It should be noted that

1: Initialize v 0 ∈ M and give a pre-defined threshold ε. 
Ns  
M−LP
2: repeat M = vH v = |kii |2 + |kmn |2 + |ku |2 .
3: Compute the Euclidean gradient ∇f (v k ) using (28); i=1 {m,n}∈E u=1
4: Compute the Riemannian gradient ∇M f (v k ) with (27); (37)
5: Choose an Armijo step size k ( [44], see section 4.2.2);
6: Update v̄ k = v k − k ∇M f (v k ); Substituting (36) into the objective function of (23), we have
7: Update v k+1 = R(v̄ k ) using (30);

Ns
8: until the gap of the objective function values between two log2 (1 + ai |v H pii |2 )
iterations is smaller than ε . i=1
9: Obtain the optimal solution v  .     
Ns
 Ns
= log2 1 + ai  kii p̄ii + kmn p̄mn
i=1 i=1 {m,n}∈E

 H 2 
B. Discussions M−LP

ii 
Next, we discuss for a sufficiently large M , why the + ku cu p 
constraint |di,j | < τ, ∀i = j in (21) can be ignored and the u=1

Ns  
solution to (22) yields small values of {dij , ∀i = j}. ai
= log2 1 + |kii |2 . (38)
Define k√ 2πd λ , f (θ, η)  cos(η) sin(θ), g(η)  sin(η), M
i=1
and p̄ij  M pij . We have
√ Therefore (23) can be re-expressed as
p̄ij  M (a∗TLIS (θit , ηit )  aRLIS (θjr , ηjr ))
1 
Ns
ai
= √ [1 . . . ejk((m1 −1)(Δfij )+(m2 −1)(Δgij ) max log2 (1 + |kii |2 )
M {kii ,kmn ,ku }
i=1
M
. . . ejk((My −1)(Δfij )+(Mz −1)(Δgij ) ]T , (31) 
Ns
s.t. |kii |2 ≤ M,
where Δfij  f (θjr , ηjr )−f (θit , ηit ) and Δgij  g(ηjr )−g(ηit ).
i=1
Therefore, we can calculate |(p̄ij )H p̄mn | as
v = [ejφ1 · · · ejφM ]H ,
|(p̄ij )H p̄mn | Ns  
M−LP
 
 My Mz  v= kii p̄ +
ii
kmn p̄ +
mn
ku cu . (39)
1    
jk((m1 −1)(ΔFijmn )+(m2 −1)(ΔGijmn ) 
= e i=1 {m,n}∈E u=1

M m =1 m =1 
1 2

      If no unit modulus constraint is placed on the phase shift vector
1  sin 12 kMy ΔFijmn sin 12 kMz ΔGijmn  v, then it is clear

that the objective function value of (39) is

=       , (32) 2
M  sin 12 kΔFijmn sin 12 kΔGijmn  maximized when N i=1 |kii | = M , which means that kmn =
s

0, ∀{m, n} ∈ E and ku = 0. As a result, we have

where ΔFijmn  Δfmn −Δfij and ΔGijmn  Δgmn −Δgij . N H
Due to the random distribution of AoA and AoD parameters, s

we have ΔFijmn = 0 or ΔGijmn = 0 for any {m, n} = {i, j} dmn = (v ) p =

 H mn
kii p̄ ii
pmn = 0 ∀{m, n} ∈ E
with probability one. Thus we have i=1
(40)
lim | (p̄ij )H p̄mn |= 0, ∀m = i  n = j, (33)
M→∞
Hence the solution to (39) or (23), v  , yields exactly zero
where  denotes the OR logic operation which means either off-diagonal entries of D. With the unit modulus constraint

N
m = i or n = j is true. Also, it is easy to verify that v = [ejφ1 · · · ejφM ]H , i=1 s
|kii |2 will not be exactly equal

Ns
(p̄ij )H p̄ij = 1. to M but the solution will force i=1 |kii |2 to be close to M
Define A  {p̄ii : i = 1, . . . , Ns } and B  {p̄mn : as much as possible in order to obtain a maximum objective
{m, n} ∈ E}, where E  U − I, in which function value. As a result, the values of {|kmn |}, ∀{m, n} ∈
U  {{m, n} : m ∈ {1, . . . , L}, n ∈ {1, . . . , P }} (34) E are small relative to {|kii |, i = 1, . . . , Ns }. Note that we
have
I  {{i, i} : i ∈ {1, . . . , Ns }} (35)
dpq = (v  )H ppq
Let C  {cu }M−LP
u=1 denote an orthonormal set whose vectors ⎛ ⎞H
are orthogonal to those vectors in A ∪ B. Clearly, the set of Ns  
M−LP
vectors in D  A ∪ B ∪ C form an orthonormal basis of CM . =⎝ kii p̄ii + kmn p̄mn + ku cu ⎠ ppq
Hence, any vector in CM can be expressed as a sum of the i=1 {m,n}∈E u=1
basis vectors scaled ⎧k
⎪
⎨√
pq
q = p

Ns  
M−LP
= M (41)
v= kii p̄ii + kmn p̄mn + ku cu . (36) ⎪
⎩√
kpp
q=p
i=1 {m,n}∈E u=1 M

Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
1058 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 2, FEBRUARY 2021

Therefore the optimal solution to (23) yields an approximately Algorithm 2 Proposed Algorithm for Joint Transceiver and
diagonal matrix D whose off-diagonal entries {dij , i = j} are LIS Design
small relative to its diagonal entries {dii , i = 1, . . . , Ns }. Input: R, G, Ns , Rt , Rr , ρ, σ 2 ;
First step
C. Ordering of {βi } and {αi } 1: Optimizing Φ by Algorithm 1 developed in Subsec-


Note that we assumed a decreasing order of the path gains tion III-A;
{βi } and {αi } earlier in this article, i.e. |β1 | ≥ |β2 | ≥ . . . ≥ Second step
|βL | and |α1 | ≥ |α2 | ≥ · · · ≥ |αP |. Nevertheless, from (22), 2: Compute the optimal precoding and combining matrices,
we can see that different orders of {βi } and {αi } may lead i.e. F opt and W opt , using (12)
to different solutions and thus different performance, and it 3: Obtain hybrid precoding and combining matrices, i.e. F BB ,
is unclear whether arranging the path gains {βi } and {αi } F RF W BB and W RF , via the manifold-based method
in a decreasing order is a good choice. To gain an insight discussed in Section IV;
into this problem, let us examine the optimization (39) which Output: Φ , F BB , F RF , W BB , and W RF .
is a variant of (22). As discussed above, if we ignore the
unit-modulus constraint placed on v, (39) can be simplified as
a conventional power allocation problem: IV. H YBRID P RECODING /C OMBINING D ESIGN

Ns
After obtaining the passive beamforming matrix Φ ,
max log2 (1 + ai |k̃ii |2 )
{k̃ii } we conduct the SVD of H  , i.e. H  = RΦ G =
i=1
U  Σ (V  )H and the optimal precoder/combiner can be

Ns
obtained via (12). As discussed earlier in this article, we search
s.t. |k̃ii |2 ≤ 1, (42)
i=1
for an analog precoding (combining) matrix F RF (W RF ) and
a baseband precoding (combining) matrix F BB (W BB ) to
where k̃ii  √kM ii
. Recall that ai = Nsρσ2 |αi βi |2 , where approximate the optimal precoder (combiner) F opt (W opt ).
 
αi = Nt M
α̃ , and β = Nr M In the following, we focus our discussion on the hybrid
P i i L β̃i . Thanks to the large
dimensions of Nt , Nr , and M , it can be expected that the precoding design as the extension to the hybrid combining
effective signal-to-noise ratios (SNRs) {ai } have large values design is similar. Such a problem can be formulated as [38]
even the nominal SNR σρ2 is small. It is well-known that in min F opt − F RF F BB 2F , (47)
the high SNR regime, it is approximately optimal to equally w∈M2 ,F BB
allocate power to different data streams. Therefore the question
where w  vec(F RF ) and the search space is a complex circle
is how to arrange the orders of {βi } and {αi } such that

Ns manifold defined as M2  S Nt Rt . Specifically, we adopt the

i=1 log2 (1 + ai /Ns ) can be maximized. fast manifold-based optimization method in [38] to solve (47),
Let {τl }Ll=1 be a permutation of the set {1, . . . , L} and where we optimize F RF and F BB in an alternating manner. The
{πp }Pp=1 a permutation of the set {1, . . . , P }. We can write
detailed procedure can be found in [38] and thus omitted here.

Ns 
Ns It should be noted that the method is guaranteed to converge to
S log2 (1 + ai /Ns ) = log2 (1 + e|απk |2 |βτk |2 ) a critical point [38], [44] and has a computational complexity
i=1 k=1 at the order of O(Nt Rt Ns L2 ), where L2 is the number of
Ns 
Ns  
2 1 2
iterations required to converge.
= log2 (|απk | ) + log2 + e|βτk |
|απk |2
k=1 k=1
V. S UMMARY AND D ISCUSSIONS
 S1 + S2 . (43)
First of all, we would like to clarify that different from
where e  N 2ρσ2 . We see that S1 is a constant once {απk }N s
k=1 other state-of-the-art methods [24], [25], our proposed method
s
is determined. Hence, we focus on the value of S2 . Note that is designed specifically for mmWave systems. It relies on
Ns 
  the sparse scattering structure of mmWave channels to obtain
1 2
2 S2 = + e|β τ | . (44) a good approximation of the SVD of the effective channel.
|απk |2 k
k=1 On the other hand, as will be shown in the subsequent
Without loss of generality, we assume analysis and simulation results, utilizing the inherent sparse
1 1 1 structure of mmWave channels enables us to achieve a better
≤ ≤ ··· ≤ . (45) tradeoff between the spectral efficiency performance and the
|απ1 |2 |απ2 |2 |απNs |2
computational complexity.
Clearly, according to the dual rearrangement inequality [46],
S2 attains its maximum if and only if
A. Computational Complexity Analysis
|βτ1 |2 ≥ |βτ2 |2 ≥ · · · ≥ |βτNs |2 . (46)
For clarity, the proposed algorithm is summarized
This means that similar to {απk }N k=1 , {|βτk |}k=1 should also
s Ns
in Algorithm 2. Specifically, our proposed algorithm consists
be arranged in a decreasing order. This explains why we of two steps. The first step aims to obtain LIS’s reflection
arrange the path gains of {βi } and {αi } in a descending order. coefficients via solving the passive beamforming optimization

Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: JOINT TRANSCEIVER AND LIS DESIGN FOR MASSIVE MIMO mmWAVE SYSTEMS
Fig. 2. Simulation setup.
problem (23). A manifold-based method, i.e. Algorithm 1,
is developed. The dominant operation for Algorithm 1 at each
iteration is to calculate the Euclidean gradient (28), which has
a computational complexity at the order of O(M ). Note that
P ii = pii (pii )H , and thus calculating (pii )H v is enough to
obtain (28). Therefore the first step involves a computational
complexity of O(M L1 ), where L1 denotes the number of iter-
ations required to converge. The second step solves a hybrid
beamforming problem which involves calculating the optimal
precoder/combiner and searching for an analog precoding
(combining) matrix and a baseband precoding (combining)
matrix to approximate the optimal precoder (combiner). The
optimal precoder/combiner can be obtained by performing the
SVD of the effective channel, whose complexity is at the order
of O(Nr Nt min(Nr , Nt )). The proposed manifold-based algo-
rithm for finding the analog/diginal precoding (combining)
matrices, as discussed in the previous section, has a complexity
of O(Nr Rr Ns L2 +Nt Rt Ns L2 ), where L2 denotes the number
of iterations required to converge. Thus, the overall computa-
tional complexity of our proposed algorithm is at the order of
O(M L1 + Nr Nt min(Nr , Nt ) + Nr Rr Ns L2 + Nt Rt Ns L2 ).
Note that there is no alternating process between these
two steps. We only need to execute each step once for our
proposed algorithm. Since each step of our proposed algorithm
is ensured to converge, the convergence of our proposed
algorithm is guaranteed.
B. Computational Complexity Comparison
In this subsection, we compare our proposed method with
some state-of-the-art methods developed for spectral efficiency
optimization for IRS-aided MIMO systems. Note that these
methods are not specially designed to mmWave systems
with hybrid precoding/combining structures. To apply them
to our problem, a similar two-step procedure is required.
Let us first consider the alternating optimization (AO)-based
algorithm [25]. In the first step, the AO-based algorithm [25]
is used to optimize the LIS’s reflection coefficients Φ
and the transmit covariance matrix Q. The first step
involves a computational complexity at the order of
O(Nr Nt (M + min(Nr , Nt ))L + ((3Nr3 + 2Nr2 Nt + Nt2 )M +
Nr Nt min(Nr , Nt ))I), where L is the number of realizations
Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
1059
TABLE I
C OMPUTATIONAL C OMPLEXITY C OMPARISON
for initialization and I denotes the number of outer iter-
ations [25]. Note that the first step requires to calculate
the SVD of the effective channel. Hence the optimal pre-
coder/combiner can be directly obtained without involving
additional computational complexity. In the second step,
similar to our work, we resort to the manifold-based algo-
rithm to find the hybrid precoder and combiner to approxi-
mate the optimal precoder/combiner, which involves a same
computational complexity at the order of O(Nr Rr Ns L2 +
Nt Rt Ns L2 ). Therefore the overall computational complex-
ity of the AO-based algorithm [25] is at the order of
O(Nr Nt (M + min(Nr , Nt ))L + ((3Nr3 + 2Nr2 Nt + Nt2 )M +
Nr Nt min(Nr , Nt ))I + Nr Rr Ns L2 + Nt Rt Ns L2 ). Note that
although both the AO-based algorithm [25] and our proposed
algorithm have a computational complexity scaling linearly
with M , our proposed algorithm achieves a substantial com-
plexity reduction as compared with [25]. To see this, suppose
Nt ≥ Nr and Nr  min{Rr , Rt } ≥ Ns , which is a reason-
able setup for practical systems. In this case, the complexity of
our proposed method is dominated by O(M L1 +Nr2 Nt ). As a
comparison, the computational complexity of the AO-based
algorithm [25] is dominated by O(2Nr2 Nt M I). Considering
the fact that L1 1 is much smaller than 2Nr2 Nt I, our proposed
method has a much lower computational complexity than the
AO-based method [25].
The complexity of other existing methods can be similarly
analyzed. Specifically, the overall computational complexity of
the WMMSE-based method in [27] is at the order of O((Nr3 +
Ns3 + max(Nt3 , Nt2 Nr ) + (M 3 + T M 2 ))I2 + Nr Rr Ns L2 +
Nt Rt Ns L2 ), where I2 (T ) denotes the outer (inner) iterations.
On the other hand, the overall computational complexity of the
sum-path-gain maximization (SPGM)-based method in [24] is
at the order of O(M 3 L3 +Nr Nt min(Nr , Nt )+Nr Rr Ns L2 +
Nt Rt Ns L2 ), where L3 denotes the number of iterations.
Assuming M ≥ Nt > Nr > Ns , the dominant compu-
tational complexity of respective algorithms is summarized
in Table I. We see that the computational complexity of
both [27] and [24] scales cubically with the number of
reflecting elements M .
VI. S IMULATION R ESULTS
In this section, we evaluate the performance of our pro-
posed joint transceiver and LIS design method. The proposed
method is based on the approximation of the truncated SVD
of the effective (cascade) channel, and thus is referred to
as Truncated-SVD-based beamforming (T-SVD-BF).2 In our
1 Simulation results show that our proposed manifold-based algorithm con-
verges within only a few, say, ten iterations.
2 Codes are available at http://www.junfang-uestc.net/codes/T-SVD-BF.rar
1060 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 2, FEBRUARY 2021

simulations, we adopt a three-dimensional setup as shown

in Fig. 2. The coordinates of the BS, the LIS and the UE
are respectively given as (xBS , 0, zBS ), (0, yLIS , zLIS ), and
(xUE , yUE , zUE ), where we set xBS = 2m, zBS = 10m,
yLIS = 148m, zLIS = 10m, xUE = 5m, yUE = 150m, and
zUE = 1.8m. The distance between the BS (LIS) and the
LIS (UE) can be easily calculated and given as 148m (9.8m).
Slightly different from our previous definition, the effective
channel is modeled as follows by considering the transmitter
and receiver antenna gains [47]

H eff = Gt Gr RΦG (48)

where Gt and Gr represent the transmitter and receiver

antenna gains, respectively. The BS-LIS channel G and the
LIS-UE channel R are modeled as follows
  Fig. 3. Spectral efficiency versus transmit power ρ.
Nt M
G= α̃1 aRLIS (θ1r , η1r )aH t
TBS (γ1 )
P
P 
+ α̃i aRLIS (θi , ηi )aTBS (γi ) ,
r r H t
(49)
i=2 
M Nr
R= β̃1 aRUE (γ1r )aH t t
TLIS (θ1 , η1 )
L

L 
+ β̃i aRUE (γi )aTLIS (θi , ηi ) ,
r H t t
(50)
i=2

where α̃1 (β̃1 ) ∼ CN (0, 10−0.1κ) denotes the complex gain of

the LOS path, κ is the path loss given by [37]
˜ +ξ
κ = a + 10b log10 (d) (51)

in which d˜ denotes the distance between the transmitter and

the receiver, and ξ ∼ N (0, σξ2 ). The values of a, b σξ
are set to be a = 61.4, b = 2, and σξ = 5.8dB as
suggested by LOS real-world channel measurements [37],
α̃i (β˜i ) ∼ CN (0, 10−0.1(κ+μ) ) stands for the complex gain of
the associated NLOS path, and μ is the Rician factor [26], [48].
Unless specified otherwise, we assume that the LIS employs
a UPA with M = My × Mz = 16 × 16, the BS and the
UE adopt ULAs with Nt = Nr = 64. Other parameters
are set as follows: Rt = Rr = 6, Ns = 4, L = P = 7,
μ = 10, Gt = 24.5dBi, Gr = 0dBi. The carrier frequency is
set to 28GHz, the bandwidth is set to 251.1886MHz and thus
the noise power is σ 2 = −174 + 10 log10 B = −90dBm.
We compare our proposed method with the following
three state-of-the-art algorithms, namely, the sum-path-gain
maximization (SPGM) method which aims to maximize
the Frobenius-norm of the effective channel H eff [24], the
AO-based algorithm [25], and the WMMSE-based algo-
rithm [27]. Note that the WMMSE-based method is designed
for multicell-multiuser scenarios. But it can easily adapted to
the point-to-point scenario considered in this article.
A. Perfect Channel State Information
In this section, we assume perfect channel state information
(CSI) is available at the BS. In Fig. 3, we plot the spectral
efficiency of respective algorithms versus the transmit power ρ,
Fig. 4. Convergence behavior of the proposed algorithm for solving (23).
where “digital” and “hybrid” respectively represent the spec-
tral efficiency achieved by fully digital precoding/combining
matrices and hybrid precoding/combining matrices. Here
hybrid precoding/combining matrices are obtained via the
manifold optimization-based method developed in Section IV.
We observe that for all four algorithms, the performance of
hybrid precoding/combining is very close to that of fully
digital precoding/combining, which corroborates the claim that
hybrid beamforming/combining with a small number of RF
chains can asymptotically approach the performance of fully
digital beamforming/combining for a sufficiently large number
of transceiver antennas. In addition, we see that our proposed
method T-SVD-BF presents a clear performance advantage
over the SPGM method and the WMMSE-based method.
Particularly, the substantial improvement over the SPGM can
be attributed to the fact that our proposed method yields
a well-conditioned effective channel matrix with a smaller
condition number, thus creating a more amiable propagation
environment for data transmission. We also note that our
proposed method achieves performance close to the AO-based
method. The convergence behavior of our proposed manifold-
based algorithm used to solve the passive beamforming design
problem (23) is depicted in Fig. 4, where we set ρ = 30dBm.

Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: JOINT TRANSCEIVER AND LIS DESIGN FOR MASSIVE MIMO mmWAVE SYSTEMS 1061

Fig. 5. Spectral efficiency and average run time versus number of LIS elements, M .

It can be seen that the proposed manifold-based algorithm has

a relatively fast convergence rate and attains the maximum
objective function value within only ten iterations.
Fig. 5 depicts the spectral efficiency and average run time
of respective algorithms versus the number of LIS elements,
where the transmit power ρ is set to 50dBm, and to vary
M , we fix My = 16 and increase Mz . All algorithms are
executed on a 2.90GHz Intel Core i7 PC with 64GB RAM.
We see that for all four algorithms, the spectral efficiency
increases as the number of passive elements increases. For
this setup, the SPGM and the WMMSE methods achieve
similar performance. Our proposed method outperforms the
SPGM and the WMMSE by a big margin, and the performance
gap becomes more pronounced as M increases. It is also
observed that the spectral efficiency of our proposed method is
slightly lower (by about 4% spectral efficiency loss) than that
of the AO-based method. Nevertheless, our proposed method
is much more computationally efficient than the AO-based
method, as reported in Fig. 5 (b). For instance, the average
run time required by our proposed method T-SVD-BF is
0.1988 second when M = 192, while it takes the AO-based
method about 26.7357 seconds to yield a satisfactory solution.
Such a substantial complexity reduction makes our proposed
method a more practical choice for LIS-assisted mmWave
communications where a large number of antennas as well as a
large number of reflecting elements are likely to be employed
to compensate the severe path loss. In addition, it is observed
that our proposed method outperforms the WMMSE-based
method in terms of both spectral efficiency and computational
complexity.
In Fig. 6, we plot the spectral efficiency versus the number
of data streams, where the transmit power ρ is set to 30dBm
and the number of passive elements is set to M = 256.
It is observed that the spectral efficiency of T-SVD-BF,
AO and WMMSE improves rapidly as Ns increases, while
the performance improvement of SPGM is quite limited.
Again, we observe that T-SVD-BF achieves a substantial
computational complexity reduction, at the expense of mild
Fig. 6. Spectral efficiency versus the number of data streams Ns .
performance degradation compared to the AO-based method.
The average run time of our proposed method is 0.5637 second
when Ns = 3, while it takes the AO-based method about
35 seconds.
To gain an insight into how the LIS helps realize a favor-
able propagation environment, we use the truncated condition
number as a metric to evaluate the capability of respective
algorithms in reconfiguring the wireless channel. As is well
known, the channel matrix condition number is regarded as
an auxiliary metric for measuring how favorable the channel
Σ21 (1,1)
is. The truncated condition number is defined as Σ2 (N s ,Ns )
,
1
where Σ1 (i, i) denotes the ith largest singular value of the
effective channel H eff . Fig. 7 shows the truncated condition
number versus the number of passive reflection elements.
We observe that the truncated condition number of SPGM
grows as M increases, whereas the truncated condition num-
bers of the proposed method and the AO-based method
are small and remain almost unchanged with an increasing
M . This result indicates that our proposed method and the
AO-based method are superior to SPGM in building a more

Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
1062
Fig. 7. Truncated condition number versus the number of LIS elements M .
Fig. 8. Spectral efficiency versus the estimation error β.
favorable wireless channel. It also explains why SPGM does
not gain much with the increase of the number of passive ele-
ments. An interesting observation is that the WMMSE-based
method performs worse than our proposed method, despite
of the fact that it has a smaller truncated condition number.
The reason lies in that although the WMMSE-based method
obtains more balanced singular values, the singular values are
generally smaller (in terms of magnitude) than those singular
values obtained by our proposed method, which prevents the
WMMSE method from achieving a higher spectral efficiency.
B. Imperfect Channel State Information
In practice, due to channel estimation errors, perfect knowl-
edge of channel state information (CSI) is usually unavail-
able. It is therefore important to evaluate the performance
of respective algorithms with inaccurate knowledge of AoAs
and AoDs. To this end, let δ  θ − θ̂ denote the estimation
error of the AoA/AOD parameter, where θ denotes the true
AoA (AoD) and θ̂ is the estimated AoA (AoD). Note that
our channel model consists of a set of AoA/AoD parameters,
and we assume that the estimation errors for these AoA/AoD
Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 2, FEBRUARY 2021
parameter are modeled as independent and identically
distributed (i.i.d.) random variables following a same uniform
distribution [49]:
1
, if − β ≤ δ ≤ β
f (δ) = 2β (52)
0, otherwise
In our simulations, we set Nt = Nr = 64, M = My × Mz =
16 × 16 = 256, L = P = 7, and ρ = 50dBm. Fig. 8 depicts
the spectral efficiency versus the estimation error β, where
β varies from 0 to 2◦ , as suggested in [49]. We see that our
proposed method behaves similarly as the AO-based method:
both methods suffer nearly a same amount of performance loss
as the estimation error increases. This result indicates that our
proposed method does not exhibit higher sensitivity to inac-
curate CSI than other methods. Moreover, it can be observed
that our proposed method still presents a clear performance
advantage over the SPGM and the WMMSE-based methods
in the presence of channel estimation errors.
VII. C ONCLUSION
In this article, we considered an LIS-assisted downlink
mmWave MIMO system with hybrid precoding/combining.
The objective is to maximize the spectral efficiency by jointly
optimizing the passive beamforming matrix at the LIS and
the hybrid precoder (combiner) at the BS (UE). To tackle this
non-convex problem, we developed a manifold optimization
(MO)-based algorithm by exploiting the inherent structure of
the effective (cascade) mmWave channel. Simulation results
showed that by carefully designing phase shift parameters at
the LIS, our proposed method can help create a favorable prop-
agation environment with a small channel matrix condition
number. Besides, it can achieve a performance comparable to
those of state-of-the-art algorithms, while at a much lower
computational complexity.
R EFERENCES
[1] T. S. Rappaport, J. N. Murdock, and F. Gutierrez, “State of the art in
60-GHz integrated circuits and systems for wireless communications,”
Proc. IEEE, vol. 99, no. 8, pp. 1390–1436, Aug. 2011.
[2] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-wave cellular
wireless networks: Potentials and challenges,” Proc. IEEE, vol. 102,
no. 3, pp. 366–385, Mar. 2014.
[3] A. Ghosh et al., “Millimeter-wave enhanced local area systems: A high-
data-rate approach for future wireless networks,” IEEE J. Sel. Areas
Commun., vol. 32, no. 6, pp. 1152–1163, Jun. 2014.
[4] S. Sun, T. S. Rappaport, M. Shafi, P. Tang, J. Zhang, and P. J. Smith,
“Propagation models and performance evaluation for 5G millimeter-
wave bands,” IEEE Trans. Veh. Technol., vol. 67, no. 9, pp. 8422–8439,
Sep. 2018.
[5] A. Alkhateeb, J. Mo, N. Gonzalez-Prelcic, and R. W. Heath, Jr., “MIMO
precoding and combining solutions for millimeter-wave systems,” IEEE
Commun. Mag., vol. 52, no. 12, pp. 122–131, Dec. 2014.
[6] A. L. Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino,
“Millimeter-wave massive MIMO: The next wireless revolution?” IEEE
Commun. Mag., vol. 52, no. 9, pp. 56–62, Sep. 2014.
[7] O. Abari, “Enabling high-quality untethered virtual reality,” in Proc. 1st
ACM Workshop Millim.-Wave Netw. Sens. Syst. (mmNets), Boston, MA,
USA: USENIX Association, Mar. 2017, pp. 531–544.
[8] G. R. MacCartney, T. S. Rappaport, and S. Rangan, “Rapid fading
due to human blockage in pedestrian crowds at 5G millimeter-wave
frequencies,” in Proc. IEEE Global Commun. Conf. (GLOBECOM),
Dec. 2017, pp. 1–7.
WANG et al.: JOINT TRANSCEIVER AND LIS DESIGN FOR MASSIVE MIMO mmWAVE SYSTEMS
[9] P. Wang, J. Fang, X. Yuan, Z. Chen, and H. Li, “Intelligent reflect-
ing surface-assisted millimeter wave communications: Joint active
and passive precoding design,” IEEE Trans. Veh. Technol., to be
published.
[10] X. Tan, Z. Sun, D. Koutsonikolas, and J. M. Jornet, “Enabling indoor
mobile millimeter-wave networks based on smart reflect-arrays,” in
Proc. IEEE Conf. Comput. Commun. (INFOCOM), Honolulu, HI, USA,
Apr. 2018, pp. 270–278.
[11] N. S. Perovic, M. D. Renzo, and M. F. Flanagan, “Channel capacity
optimization using reconfigurable intelligent surfaces in indoor mmWave
environments,” in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2020,
pp. 1–7.
[12] X. Yang, C.-K. Wen, and S. Jin, “MIMO detection for reconfigurable
intelligent surface-assisted millimeter wave systems,” IEEE J. Sel. Areas
Commun., vol. 38, no. 8, pp. 1777–1792, Aug. 2020.
[13] C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis,
and I. Akyildiz, “A new wireless communication paradigm through
software-controlled metasurfaces,” IEEE Commun. Mag., vol. 56, no. 9,
pp. 162–169, Sep. 2018.
[14] Y.-C. Liang, R. Long, Q. Zhang, J. Chen, H. V. Cheng, and H. Guo,
“Large intelligent surface/antennas (LISA): Making reflective radios
smart,” J. Commun. Inf. Netw., vol. 4, no. 2, pp. 40–50, Jun. 2019.
[15] S. Hu, F. Rusek, and O. Edfors, “Beyond massive MIMO: The potential
of data transmission with large intelligent surfaces,” IEEE Trans. Signal
Process., vol. 66, no. 10, pp. 2746–2758, May 2018.
[16] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini, and
R. Zhang, “Wireless communications through reconfigurable intelligent
surfaces,” IEEE Access, vol. 7, pp. 116753–116773, 2019.
[17] Q. Wu and R. Zhang, “Towards smart and reconfigurable environment:
Intelligent reflecting surface aided wireless network,” IEEE Commun.
Mag., vol. 58, no. 1, pp. 106–112, Jan. 2020.
[18] Y. Han, W. Tang, S. Jin, C.-K. Wen, and X. Ma, “Large intelli-
gent surface-assisted wireless communication exploiting statistical CSI,”
IEEE Trans. Veh. Technol., vol. 68, no. 8, pp. 8238–8242, Aug. 2019.
[19] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
network: Joint active and passive beamforming design,” in Proc. IEEE
Global Commun. Conf. (GLOBECOM), Abu Dhabi, UAE, Dec. 2018,
pp. 1–6.
[20] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in
wireless communication,” IEEE Trans. Wireless Commun., vol. 18, no. 8,
pp. 4157–4170, Aug. 2019.
[21] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
network via joint active and passive beamforming,” IEEE Trans. Wireless
Commun., vol. 18, no. 11, pp. 5394–5409, Nov. 2019.
[22] X. Li, J. Fang, F. Gao, and H. Li, “Joint active and passive beamforming
for intelligent reflecting surface-assisted massive MIMO systems,” 2019,
arXiv:1912.00728. [Online]. Available: http://arxiv.org/abs/1912.00728
[23] M. Cui, G. Zhang, and R. Zhang, “Secure wireless communication via
intelligent reflecting surface,” IEEE Wireless Commun. Lett., vol. 8,
no. 5, pp. 1410–1414, Oct. 2019.
[24] B. Ning, Z. Chen, W. Chen, and J. Fang, “Beamforming optimiza-
tion for intelligent reflecting surface assisted MIMO: A sum-path-gain
maximization approach,” IEEE Wireless Commun. Lett., vol. 9, no. 7,
pp. 1105–1109, Jul. 2020.
[25] S. Zhang and R. Zhang, “Capacity characterization for intelligent
reflecting surface aided MIMO communication,” IEEE J. Sel. Areas
Commun., vol. 38, no. 8, pp. 1823–1838, Aug. 2020.
[26] K. Ying, Z. Gao, S. Lyu, Y. Wu, H. Wang, and M.-S. Alouini,
“GMD-based hybrid beamforming for large reconfigurable intelligent
surface assisted millimeter-wave massive MIMO,” IEEE Access, vol. 8,
pp. 19530–19539, 2020.
[27] C. Pan et al., “Multicell MIMO communications relying on intelligent
reflecting surfaces,” IEEE Trans. Wireless Commun., vol. 19, no. 8,
pp. 5218–5233, Aug. 2020.
[28] C. Pan et al., “Intelligent reflecting surface aided MIMO broadcasting
for simultaneous wireless information and power transfer,” IEEE J. Sel.
Areas Commun., vol. 38, no. 8, pp. 1719–1734, Aug. 2020.
[29] D. Mishra and H. Johansson, “Channel estimation and low-complexity
beamforming design for passive intelligent surface assisted MISO wire-
less energy transfer,” in Proc. IEEE Int. Conf. Acoust., Speech Signal
Process. (ICASSP), May 2019, pp. 4659–4663.
[30] P. Wang, J. Fang, H. Duan, and H. Li, “Compressed channel estimation
for intelligent reflecting surface-assisted millimeter wave systems,” IEEE
Signal Process. Lett., vol. 27, pp. 905–909, 2020.
Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
1063
[31] J. Chen, Y.-C. Liang, H. Victor Cheng, and W. Yu, “Chan-
nel estimation for reconfigurable intelligent surface aided multi-
user MIMO systems,” 2019, arXiv:1912.03619. [Online]. Available:
http://arxiv.org/abs/1912.03619
[32] T. Lin, X. Yu, Y. Zhu, and R. Schober, “Channel estimation for intelli-
gent reflecting surface-assisted millimeter wave MIMO systems,” 2020,
arXiv:2005.04720. [Online]. Available: http://arxiv.org/abs/2005.04720
[33] X. Yu, J.-C. Shen, J. Zhang, and K. B. Letaief, “Alternating minimization
algorithms for hybrid precoding in millimeter wave MIMO systems,”
IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 485–500,
Apr. 2016.
[34] H. Sawada et al., LOS Office Channel Model Based on TSV Model,
Standard IEEE 802. 15-06-0377-02-003c, Sep. 2006.
[35] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, Jr.,
“Spatially sparse precoding in millimeter wave MIMO systems,” IEEE
Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
[36] X. Gao, L. Dai, S. Han, C.-L. I, and R. W. Heath, Jr., “Energy-efficient
hybrid analog and digital precoding for mmWave MIMO systems with
large antenna arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4,
pp. 998–1009, Apr. 2016.
[37] M. R. Akdeniz et al., “Millimeter wave channel modeling and cellular
capacity evaluation,” IEEE J. Sel. Areas Commun., vol. 32, no. 6,
pp. 1164–1179, Jun. 2014.
[38] H. Kasai, “Fast optimization algorithm on complex oblique mani-
fold for hybrid precoding in millimeter wave MIMO systems,” in
Proc. IEEE Global Conf. Signal Inf. Process. (GlobalSIP), Nov. 2018,
pp. 1266–1270.
[39] T. L. Marzetta, Fundamentals of Massive MIMO. Cambridge, U.K.:
Cambridge Univ. Press, 2016.
[40] J. Chen, “When does asymptotic orthogonality exist for very large
arrays?” in Proc. IEEE Global Commun. Conf. (GLOBECOM), Atlanta,
GA, USA, Dec. 2013, pp. 4146–4150.
[41] Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An iteratively weighted
MMSE approach to distributed sum-utility maximization for a MIMO
interfering broadcast channel,” IEEE Trans. Signal Process., vol. 59,
no. 9, pp. 4331–4340, Sep. 2011.
[42] Q. Shi and M. Hong, “Spectral efficiency optimization for millimeter
wave multiuser MIMO systems,” IEEE J. Sel. Topics Signal Process.,
vol. 12, no. 3, pp. 455–468, Jun. 2018.
[43] K. Alhujaili, V. Monga, and M. Rangaswamy, “Transmit MIMO radar
beampattern design via optimization on the complex circle manifold,”
IEEE Trans. Signal Process., vol. 67, no. 13, pp. 3561–3575, Jul. 2019.
[44] P.-A. Absil, R. Mahony, and R. Sepulchre, Optimization Algorithms on
Matrix Manifolds. Princeton, NJ, USA: Princeton Univ. Press, 2009.
[45] N. Boumal, B. Mishra, P.-A. Absil, and R. Sepulchre, “Manopt, a
MATLAB toolbox for optimization on manifolds,” J. Mach. Learn. Res.,
vol. 15, pp. 1455–1459, Aug. 2014.
[46] A. Oppenheim, “Inequalities connected with definite hermitian forms,
II,” Amer. Math. Monthly, vol. 61, no. 7P1, pp. 463–466, Aug. 1954.
[47] B. Ning, Z. Chen, W. Chen, Y. Du, and J. Fang, “Terahertz multi-
user massive MIMO with intelligent reflecting surface: Beam training
and hybrid beamforming,” 2019, arXiv:1912.11662. [Online]. Available:
http://arxiv.org/abs/1912.11662
[48] M. K. Samimi, G. R. MacCartney, S. Sun, and T. S. Rappaport, “28 GHz
millimeter-wave ultrawideband small-scale fading models in wireless
channels,” in Proc. IEEE 83rd Veh. Technol. Conf. (VTC Spring),
May 2016, pp. 1–6.
[49] C. Pradhan, A. Li, L. Zhuo, Y. Li, and B. Vucetic, “Beam misalignment
aware hybrid transceiver design in mmWave MIMO systems,” IEEE
Trans. Veh. Technol., vol. 68, no. 10, pp. 10306–10310, Oct. 2019.
Peilan Wang received the B.Sc. degree from the
University of Electronic Science and Technology
of China, Chengdu, China, in 2018, where she
is currently pursuing the Ph.D. degree. Her cur-
rent research interests include compressed sensing,
millimeter wave, reflecting intelligent surface, and
signal processing.
1064 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 2, FEBRUARY 2021
Jun Fang (Senior Member, IEEE) received the B.S.
and M.S. degrees from Xidian University, Xi’an,
China, in 1998 and 2001, respectively, and the Ph.D.
degree from the National University of Singapore,
Singapore, in 2006, all in electrical engineering.
In 2006, he was a Post-Doctoral Research Asso-
ciate with the Department of Electrical and Com-
puter Engineering, Duke University. From Janu-
ary 2007 to December 2010, he was a Research
Associate with the Department of Electrical and
Computer Engineering, Stevens Institute of Technol-
ogy. Since 2011, he has been with the University of Electronic of Science
and Technology of China. His research interests include compressed sensing
and sparse theory, massive MIMO/mmwave communications, and statistical
inference.
Dr. Fang received the IEEE Jack Neubauer Memorial Award in 2013 for the
Best Systems Paper published in the IEEE T RANSACTIONS ON V EHICULAR
T ECHNOLOGY. He serves as a Senior Associate Editor for the IEEE S IGNAL
P ROCESSING L ETTERS .
Linglong Dai (Senior Member, IEEE) received the
B.S. degree from Zhejiang University, Hangzhou,
China, in 2003, the M.S. degree (Hons.) from the
China Academy of Telecommunications Technol-
ogy, Beijing, China, in 2006, and the Ph.D. degree
(Hons.) from Tsinghua University, Beijing, in 2011.
From 2011 to 2013, he was a Post-Doctoral Research
Fellow with the Department of Electronic Engineer-
ing, Tsinghua University, where he was an Assistant
Professor from 2013 to 2016 and has been an
Associate Professor since 2016. He has coauthored
the book MmWave Massive MIMO: A Paradigm for 5G (Academic Press,
2016). He has authored or coauthored over 60 IEEE journal articles and
over 40 IEEE conference papers. He also holds 19 granted patents. His
current research interests include massive MIMO, millimeter-wave commu-
nications, THz communications, NOMA, reconfigurable intelligent surface
(RIS), and machine learning for wireless communications. He has received
five IEEE Best Paper Awards at IEEE ICC 2013, IEEE ICC 2014, IEEE ICC
2017, IEEE VTC 2017-Fall, and IEEE ICC 2018. He has also received the
Tsinghua University Outstanding Ph.D. Graduate Award in 2011, the Beijing
Excellent Doctoral Dissertation Award in 2012, the China National Excellent
Doctoral Dissertation Nomination Award in 2013, the URSI Young Scientist
Award in 2014, the IEEE T RANSACTIONS ON B ROADCASTING Best Paper
Award in 2015, the Electronics Letters Best Paper Award in 2016, the National
Natural Science Foundation of China for Outstanding Young Scholars in 2017,
the IEEE ComSoc Asia-Pacific Outstanding Young Researcher Award in 2017,
the IEEE ComSoc Asia-Pacific Outstanding Paper Award in 2018, the China
Communications Best Paper Award in 2019, and the IEEE Communications
Society Leonard G. Abraham Prize in 2020. He is also an Area Editor
of the IEEE C OMMUNICATIONS L ETTERS , and an Editor of the IEEE
T RANSACTIONS ON C OMMUNICATIONS and the IEEE T RANSACTIONS ON
V EHICULAR T ECHNOLOGY. He is particularly dedicated to reproducible
research and has made a large amount of simulation code publicly available.
Authorized licensed use limited to: Tsinghua University. Downloaded on June 07,2021 at 02:25:26 UTC from IEEE Xplore. Restrictions apply.
Hongbin Li (Fellow, IEEE) received the B.S. and
M.S. degrees from the University of Electronic Sci-
ence and Technology of China in 1991 and 1994,
respectively, and the Ph.D. degree from the Uni-
versity of Florida, Gainesville, FL, in 1999, all in
electrical engineering.
From July 1996 to May 1999, he was a Research
Assistant with the Department of Electrical and
Computer Engineering, University of Florida. Since
July 1999, he has been with the Department of Elec-
trical and Computer Engineering, Stevens Institute
of Technology, Hoboken, NJ, where he is currently the Charles and Rosanna
Batchelor Memorial Chair Professor. He was a Summer Visiting Faculty
Member at the Air Force Research Laboratory in the summers of 2003, 2004,
and 2009. His general research interests include statistical signal processing,
wireless communications, and radars.
Dr. Li is a member of Tau Beta Pi and Phi Kappa Phi. He received the Sigma
Xi Graduate Research Award from the University of Florida in 1999, the Jess
H. Davis Memorial Research Award from the Stevens Institute of Technology
in 2001, the Harvey N. Davis Teaching Award in 2003, the Outstanding
Paper Award from the IEEE AFICON Conference in 2011, the IEEE Jack
Neubauer Memorial Award from the IEEE Vehicular Technology Society
in 2013, and the Provost’s Award for Research Excellence in 2019. He has
been involved in various conference organization activities, including serving
as the General Co-Chair of the 7th IEEE Sensor Array and Multichannel
Signal Processing (SAM) Workshop, Hoboken, NJ, June 17–20, 2012. He has
been a member of the IEEE SPS Signal Processing Theory and Methods
Technical Committee (TC) and the IEEE SPS Sensor Array and Multichannel
TC, an Associate Editor of Signal Processing, the IEEE T RANSACTIONS
ON S IGNAL P ROCESSING , the IEEE S IGNAL P ROCESSING L ETTERS , and
the IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS, as well as
a Guest Editor of the IEEE J OURNAL OF S ELECTED T OPICS IN S IGNAL
P ROCESSING and EURASIP Journal on Applied Signal Processing.